lkml.org 
[lkml]   [2002]   [Jan]   [5]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [announce] [patch] ultra-scalable O(1) SMP and UP scheduler
Ingo Molnar wrote:
> yes, we can do the following: instead of sending the broadcast IPI, we can
> skip CPUs that run a RT task that has a higher priority than the one that
> just got woken up. This way we send the IPI only to CPUs that have a
> chance to actually preempt their current task for the newly woken up task.
> We do this optimization for SCHED_OTHER tasks already, behavior like this
> is not something cast into stone.
>
> this way you can mark your RT task 'uninterruptible' by giving it a high
> RT priority.

This is sort of what I was thinking, although it makes it more expensive
to determine to IPI mask. A very light weight alternative could be to
make only the highest priority RT tasks uninterruptible. If a CPU is
busy with one of these it can simply be taken out of the global group
for the duration. If a task is worried about interruptions, it can boost
its prio to max for the critical part and the return it to normal
afterwards.

> i'd also like to note that Davide's description made the broadcast IPI
> solution sound more scary than it is in fact. A broadcast IPI's handler is
> pretty lightweight (it does a single APIC ACK and returns), and even a
> pointless trip into the O(1) scheduler wont take more time than say 10-20
> microseconds (pessimistic calculation), on a typical x86 system.

I guess the worry is that the overhead isn't bounded and the per-CPU
overhead increases with the number of CPUs. If the interrupts happen
come in a burst, a running task can experience a longer interruption.

> The reason i made the IPI a broadcast in the RT case is race avoidance:
> right now our IPIs are 'inexact', ie. if the scheduling situation changes
> while they are in flight (they can take 5-10 microseconds to get delivered
> to the target) then they might hit the wrong target. In case of RT/SMP,
> this might end up us missing to run a task that should be run. This was
> the major reason why i took the broadcast IPI solution.

I see your point. I may be getting off base here, but how does the cost
of making the IPIs exact compare to the cost of using broadcast IPIs on
an n-way machine (depends on n I suppose)?

Regards,

MikaL
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

\
 
 \ /
  Last update: 2005-03-22 13:15    [W:0.076 / U:2.788 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site